History rows now carry a source label: manual edits and merchant rules
are the user's decisions, ranked ahead of equally similar rows the
model classified itself and guaranteed slots in a full history window.
Without the distinction, precedent fed the model its own uncorrected
answers as majority evidence, so a correction never won against the
rows it was meant to fix. Both system prompts state that user entries
outrank ai entries. Alias write-back on manual merchant links and the
per-merchant usual category already learned locally; this closes the
loop for categories and tags.
Analyse now classifies up to ten same-kind transactions per provider
request: the registry and history travel once per batch, so a
thousand-row backfill costs about a hundred paced requests instead of a
thousand. The answer schema appears once — an array item carrying an
enum-bound ref — because providers meter strict schemas by token cost:
duplicating registry enums per row, or bounding arrays with
minItems/maxItems that Gemini expands per element, rejects real
registries with a bare HTTP 400. Row count, duplicate refs, duplicate
tags and taxonomy bounds are all enforced server-side instead, and a
request still rejected outright halves until accepted, remembering the
working size for the run. Batch requests scale the HTTP budget by row
count, chunk failures cannot abort a run whose later rows succeeded,
and rows resolved against one snapshot share one minted merchant.
Measured on a real 165-row month over a zero-data-retention route:
165 analysed, 152 proposals, 0 errors, 17 requests, under 8 minutes.
Fresh installs default to google/gemini-3.8-flash, the model that
demonstrably honors strict structured outputs over a ZDR route. Preview
changes now carry counterparty, amount and currency, and the review
list shows the amount with a counterparty fallback for banks that leave
descriptions empty.
Azure is the only zero-data-retention route for the gpt-5.6 family, so
its capacity 429s arrive frequently and OpenRouter forwards them inside
an HTTP 200 envelope. Those bypassed the rate controller entirely: a
paced run kept sending a request every three seconds into a throttled
endpoint, failing row by row. An in-envelope 429 now records the same
escalating cooldown as a transport 429, so later acquisitions fail fast
until the deadline passes.
Strict structured-output mode rejects uniqueItems, so every request to a
gpt-5.6-family zero-data-retention endpoint failed with HTTP 400 behind
a generic error; duplicates were already rejected server-side, so the
keyword leaves the wire schemas, pinned by a strict-keyword allowlist
test built from the ledger that hit this.
The bare-BIC redaction pattern deleted every 8- and 11-letter word —
Openbank, BAUMARKT, RACETRACKER — blinding the model to the payee it
was asked to classify and tripping the unsafe-merchant check on honest
answers. BICs now die only labeled or attached to their IBAN, account
labels join the redaction secrets, an identifier-shaped merchant name
degrades to a merchant-less proposal instead of failing the row, and a
provider error inside an HTTP 200 envelope is reported as such (numeric
code only) instead of as envelope corruption.
The gpt-5.6 family's zero-data-retention endpoints declare
max_completion_tokens, so sending max_tokens under require_parameters
excluded every ZDR route and returned HTTP 404 for the whole family.
The cap is retired: the strict schema, the finish_reason check and the
64 KiB read cap already bound the response.
The model fields now offer the provider's public ZDR catalog filtered
by the exact conditions completions are routed under (live endpoint,
strict structured outputs), fetched server-side, cached for an hour,
and served at GET /api/models; the inputs stay free text so an unlisted
model remains usable when the catalog is unreachable.
A low-confidence answer was discarded inside Classify, so Analyse showed
the row as unchanged instead of a reviewable suggestion; the fallback
decision moves to the import path, which keeps the merchant link and the
recorded confidence. ApplyPreview now rebases onto the current journal:
unrelated commits during a minutes-long paced run no longer invalidate
the review, only an edit to a selected transaction itself conflicts, and
applied changes are pruned so the rest stay appliable.
An account now has a kind, and an investment account holds positions as well as
cash. A broker row is not a new entity: it is a bank fact with an optional
position leg, so deduplication, the journal, fact immutability, the DuckDB
projection and the transactions view carry it unchanged. Facts.Amount stays the
cash leg and is zero on the rows that move only a position.
Scalable Capital exports are recognized locally as a fourth format, read by
their own parser because a column mapping cannot describe them: the amount
column is settled cash on a cash row, a gross to be netted on a trade, and a
position valuation that must never touch cash on a corporate action or a depot
transfer. A cash amount is already net of the tax the broker withheld or
refunded, so that tax is recorded on the fact and never subtracted a second
time; treating a corporate action's valuation as money conjures cash, and a
depot switch would do it once per instrument. The share column is signed only
for those two types, so buys and sells take their direction from the type. Every
security row is checked against shares times price at 128-bit width, because a
lost decimal separator survives every other check. An unknown status, type or
assetType, a foreign currency, a missing ISIN, or one failed check rejects the
whole file with the record number.
Instruments live in instruments.finance, keyed by ISIN with an ID derived from
it, so re-importing never registers a security twice. One ISIN appears under
several broker descriptions over the years and sometimes under the ISIN itself:
the most recent real description names it, and an import never renames one that
already exists. A broker also reuses a single reference across every leg of one
event, so transaction identity includes the event and its instrument.
domain.Fallback returns kind "investment" for any fact carrying a position leg,
so no broker row reaches the sign-based branch. That single rule is what stops
an unmatched deposit from counting as income and a broker fee from counting as
household spending; the monthly PRIME fee and its matching credit now cancel in
clearing:investments with no configuration at all. Investment rows are excluded
from spending analytics, from bulk reclassification and from the model, exactly
as transfers are.
Equal competing transfers are paired instead of skipped. Every connected
component of the candidate graph is a complete bipartite graph between two fixed
accounts at one amount and currency, so every pairing produces the same
accounts, kinds and postings and only the displayed counterpart differs.
Refusing to choose was the expensive option: both legs fell through to the
sign-based fallback and appeared as spending and income that never happened.
Pairing follows the nearest booking date, then the transaction ID, so iteration
order decides nothing. POST /api/transactions/{id}/transfer rewrites the old and
the new pair in one commit, because reciprocity is validated and a half-applied
link is an invalid dataset, and the matcher now skips any record classified
manually so a hand-made link or unlink outlives the next import.
Wealth reports each account's cash and positions from the journal rather than
the index, with named checks - row arithmetic, cash never negative, holdings
never negative - because it exists to be compared against the figures a broker
shows on its own screen. A negative holding means the imported history is
partial. Share counts are exact to eight places; a reinvested distribution
quoted to six is rounded to money's four and the residue is reported rather than
hidden. Market prices, market value, net worth over time, FIFO lot accounting,
realised gains and currency conversion are deliberately absent.
Classification preferences gains "Classify newly imported transactions with
AI", stored as classify_on_import in config.toml and on by default, so existing
configurations keep their behaviour. It covers CSV imports and bank
synchronization alike.
With it off, no import path contacts the provider: classification falls to the
new provider-free rules path, where an opted-in merchant rule still applies its
category and tags, an alias match still attaches its merchant, and everything
else arrives on the editable fallback without a provenance error that would
suggest the provider had failed. Analyse remains available on demand.
CSV import is now mapping-driven: N26, ING (metadata preamble, Windows-1252,
German decimals) and Kontist exports are recognized locally, and any other
layout can have its columns proposed by the configured model from a sample in
which letters are replaced by x and digits by 0. Proposals are untrusted: every
column must name a supplied header, money must come from one signed column or
one debit/credit pair, and formats must be from a closed list.
Uploading no longer imports. /api/import is replaced by prepare/confirm/cancel:
prepare parses, deduplicates and previews the exact facts, and only confirming
at the reviewed revision writes them. ING and AI-mapped facts carry no
transaction reference, because repeating SEPA mandate references must never
become a transaction identity.
Lars Nolden